An ensemble of convolutional neural networks for pattern recognition?

Question

Neural networks

Is it advisable to use an ensemble of convolutional neural networks for pattern recognition problems?
According to the following English-language article and related presentation , the use of an ensemble of convolutional neural networks leads to a significant reduction in errors. Shown on the example of digit pattern recognition from the MNIST database .

It is shown that with an increase in the number of models in an ensemble, recognition accuracy increases:
An ensemble algorithm can lead not only to an improvement in recognition accuracy, but also to a decrease in the time spent on training.

I could not find the year of publication of this article, as well as the degree of its author. Perhaps the use of an ensemble of convolutional neural networks is no longer relevant today or has shown its inconsistency in pattern recognition problems.
I also drew attention to the following Russian-language article dating back to 2012:

Topical issues of using convolutional neurons...

It also performs image recognition of numbers from various databases. The article and work clearly demonstrate the effectiveness of the use of CNN committees trained on bases with different styles of writing.
Here is one of the tables comparing the accuracy of recognition by committees of neural networks and the KADMOS system:

But I myself came across opposing opinions of experts who argued that the use of ensembles of convolutional neural networks is not relevant today. According to them, today the solution of pattern recognition problems is more concentrated on optimizing one neural network model than on using their ensembles.
And yet, how are things in this matter? At the moment, are there any significant advantages of using an ensemble of convolutional neural networks over the option with one neural network (without an ensemble) in pattern recognition problems or not?

Answer 1

Одна сеть - выдаёт один класс коэффициентов. Несколько - несколько.
При использовании нескольких с сумматором - конкретную ошибку в какой-либо отдельной сети установить очень сложно.
Гораздо лучше использовать подход:
"многоярусный адаптивный водопад" (с)2016, xmoonlight,
когда используя выходные данные одной сети (или нескольких параллельных), выбирается нужная ветка другой согласно какому-то условию (как правило, это абсолютный интервал выходных значений весов в текущем или предыдущем цикле в процентном соотношении относительно друг друга). В этом случае - отловить ошибку на любом этапе очень просто.